Data leakage & model-side risks

Everything you put in a prompt leaves your process — it reaches a provider, may be retained, and can resurface. Everything you log stays with you, often for years. Both paths leak the same things: emails, tokens, keys, customer data.

This module builds the redactor that guards both, then steps back to the systemic risks: system-prompt exfiltration (treat it as not secret), output filtering, and tenant isolation — which is really P56 authorization applied to a vector store.